Professional Documents
Culture Documents
The goal was to gain new knowledge and develop expertise about the fine
structure functionality of packet data traffic for the development of future mobile
data systems.
To create a model, which adequately describes the characteristics of the
individual users connection over different time scales.
A special interest in the lower levels of the time scale
Packet data traffic measurements for
- WWW service in laboratory LAN 1996 and 1999 and WLAN 1997.
- WAP over GSM data and over GPRS from a test WAP-gateway 2000-2002
The data was grouped by individual connections and analyzed based on the
protocols used.
The statistics were modeled based on events of WWW session
- Developed from the ETSI packet data model (referred Ch. 10.1)
- measured distributions are fitted to some analytic distributions
- aim is to get parameters for simulation model(s)
- intended to developing radio link protocols and radio network planning
Wired vs. Mobile Data Traffic
In fixed networks
- bandwidth is large and rapidly growing and transmission errors are rare
- most crucial elements are the centralized components like main trunks, routers
or servers
- one of the main problems is the aggregate traffic of numerous users, which
overloads these relatively few "bottlenecks"
=> the traffic should be measured from the "hot spots".
In mobile networks
- Bandwidth is quite limited and the probability of transmission errors is rather
high
- Few active users can make use of most of the traffic capacity available in a
cell
- The main "bottleneck" is the air interface at the edge of network
=> the traffic should be measured as close to the client as possible.
WAP traffic
to provide a mobile user a WWW like access to the Internet.
a HTTP-like protocol optimized to the wireless domain.
Uses TCP and IP protocols for transmission
The measurements used circuit switched GSM data and WAP protocol 1.0.
The traffic logged simultaneously from both sides of the gateway.
The effects of wireless and Internet connection and the gateway separated
already the activity during WAP-transaction < voice activity (esp. uplink)
=> advantages by multiplexing several sources to shared channels.
implementation errors generate "pseudo traffic"
the distributions have been enhanced to fit better to the measured data
no zero length delays => shift (= fixed delay) added to exponential distribution
bias (= fixed value of zero)
The Pareto distribution
is defined by
k
f x ( x) = +1 ,x k (0.1)
x
0 ,k < x
Fx ( x) = k )
1 ( ,k x (0.2)
x
k
= , 1 (0.3)
1
k 2
=2
, > 2 (0.4)
( 2) ( 1)
when < 2 the variance and when < 1 also the mean become infinite
normally the Pareto distribution is limited to area 1 < < 2
Truncating the Pareto distribution
parameter T added to compress the in principle unlimited Pareto distribution to
the practice
0 ,k < x
1 (k x )
Fx ( x) =
,k x T
1 ( k T ) (0.5)
1 ,x >T
floor ( x )
F ( x p) =
i =0
pq i where q = 1 p (0.6)
Since the mathematical distribution starts from zero to reach the aimed mean P
must be set
1
P= (0.7)
1 + mean
The developed traffic data models
The selected statistics were fitted to analytic distributions
simple model is one CDF and partial model is weighted sum of one exponential
and two truncated Pareto CDFs
models are a collection of several measurable distributions on different levels in
top-down order
the mean and variance for the measured data and the models
error value used as the measure in curve fitting
To make comparison easier each distribution for both models are presented in a
table for both 1996 and 1999 measurements. In a third table there is a comparison
of the mean and variance for the measured data and the both models. There is
also the error value, which was used as the measure in numerical curve fitting and
optimization. It is the sum of squared error between the CDF vectors for measured
data and the model.
A sample distribution for Dwww
4.1 The WWW-session interarrival time, Dwww
D www
D p c
S e s s io n
N p c
D pc
Session
Npc
T 0 b T 0 e
W S P /
W D P
T 0 a T 0 d
Page D pii D pii
T 2 b T 2 e Ni
W S P /
W T P
T 2 a T 2 d T 2 f D iud Didu
T 0 c / T 2 c
Item
W W W -
Ite m
D iuu D idd
Figure 2. The model of WAP transaction timing. Figure 3. The model of WWW-session timing.
The ETSI Non-Real Time Traffic Model
Referred later as ETSI model. Presented in D-ETR SMG-50402 v0.9.3: 4/1997,
Annex 2, and enhanced in Technical Report TR 101 112 V3.2.0 (1998-04) Annex
B, in both Paragraph 1.2.2. Traffic models / Non-real time services
A packet call
Table 1.1 Characteristics of connection-less information types (default mean values for the distributions of typical
www service)
Packet based Average number Average reading Average amount of Average Parameters for
information types of packet calls time between packets within a interarrival packet size
within a session packet calls [s] packet call [] time between distribution
1
packets [s]
WWW surfing
UDD 8 kbit/s 5 412 25 0.5 k = 81.5
UDD 32 kbit/s 5 412 25 0.125 = 1.1
UDD 64 kbit/s 5 412 25 0.0625
UDD 144 kbit/s 5 412 25 0.0277
UDD 384 kbit/s 5 412 25 0.0104
UDD 2048 kbit/s 5 412 25 0.00195
(originally
UDD 8 kbit/s) 5 12 15 0.96
1
The different interarrival times correspond to average bit rates of 8, 32, 64, 144, 384 and 2048 kbit/s.
According to the values for and k in the Pareto distribution, the average packet size n is 480 and average requested file-
size is Nd x = 25 x 480 bytes 12 kBytes. The packet size is limited to 66 666 bytes, giving a finite variance to the distri-
bution. (First the truncations effect were neglected giving n = 896 bytes and Nd x = 15 x 896 bytes 13,4 kBytes.)
The principle of dividing the model to layers like session, packet call and a
packet is very good and describes the quite closely the actual process
major drawback in the presented model are:
1. it does not take in to the consideration the direction of the packets
- measured WWW traffic has great asymmetry
- delays are different for example up to Down (~RTT) and down to up
- used protocols can differ between Uplink and Downlink
2. WWW-pages are often composed of several (on average 4.8) WWW-items which use
more than one parallel TCP-connections.
3. the systematic usage of selected statistic distributions can mask out some typical
features.
- For example the datagram (=packet) size and average interarrival time distributions.
The timing diagram presented in the figure 2. WAP transaction is there divided in
following parts:
Table 3. The average times for duration and separation for WAP-transactions, WAP-pages and
WAP-sessions.
The activity during WAP-transactions and -sessions
The activity we defined as the minimum time needed to transfer the measured IP-packets over the given
bandwidth